This article was downloaded by: [McMaster University]

On: 28 December 2014, At: 11:48

Publisher: Taylor & Francis
Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered
office: Mortimer House, 37-41 Mortimer Street, London W1T 3JH, UK

Journal of Experimental & Theoretical

Artificial Intelligence
Publication details, including instructions for authors and
subscription information:
http://www.tandfonline.com/loi/teta20

Using genetic algorithms to create

meaningful poetic text
a b c
Ruli Manurung , Graeme Ritchie & Henry Thompson
a
Faculty of Computer Science , Universitas Indonesia , Depok
16424 , Indonesia
b
Department of Computing Science , University of Aberdeen ,
Aberdeen AB24 3UE , UK
c
School of Informatics , University of Edinburgh , Edinburgh EH8
9AB , UK
Published online: 27 Jun 2011.

To cite this article: Ruli Manurung , Graeme Ritchie & Henry Thompson (2012) Using genetic
algorithms to create meaningful poetic text, Journal of Experimental & Theoretical Artificial
Intelligence, 24:1, 43-64, DOI: 10.1080/0952813X.2010.539029

To link to this article: http://dx.doi.org/10.1080/0952813X.2010.539029

PLEASE SCROLL DOWN FOR ARTICLE

Taylor & Francis makes every effort to ensure the accuracy of all the information (the
“Content”) contained in the publications on our platform. However, Taylor & Francis,
our agents, and our licensors make no representations or warranties whatsoever as to
the accuracy, completeness, or suitability for any purpose of the Content. Any opinions
and views expressed in this publication are the opinions and views of the authors,
and are not the views of or endorsed by Taylor & Francis. The accuracy of the Content
should not be relied upon and should be independently verified with primary sources
of information. Taylor and Francis shall not be liable for any losses, actions, claims,
proceedings, demands, costs, expenses, damages, and other liabilities whatsoever or
howsoever caused arising directly or indirectly in connection with, in relation to or arising
out of the use of the Content.

This article may be used for research, teaching, and private study purposes. Any
substantial or systematic reproduction, redistribution, reselling, loan, sub-licensing,
systematic supply, or distribution in any form to anyone is expressly forbidden. Terms &
 Conditions of access and use can be found at http://www.tandfonline.com/page/terms-
and-conditions
Downloaded by [McMaster University] at 11:49 28 December 2014
 Journal of Experimental & Theoretical Artificial Intelligence
Vol. 24, No. 1, March 2012, 43–64

Using genetic algorithms to create meaningful poetic text

Ruli Manurunga*, Graeme Ritchieb and Henry Thompsonc
a
Faculty of Computer Science, Universitas Indonesia, Depok 16424, Indonesia;
b
Department of Computing Science, University of Aberdeen, Aberdeen AB24 3UE, UK;
c
School of Informatics, University of Edinburgh, Edinburgh EH8 9AB, UK
(Final version received 1 November 2010)
Downloaded by [McMaster University] at 11:49 28 December 2014

This article presents a series of experiments in automatically generating poetic

texts. We confined our attention to the generation of texts which are syntactically
well-formed, meet certain pre-specified patterns of metre and broadly convey
some given meaning. Such aspects can be formally defined, thus avoiding the
complications of imagery and interpretation that are central to assessing more
free forms of verse. Our implemented system, McGONAGALL, applies the genetic
algorithm to construct such texts. It uses a sophisticated linguistic formalism to
represent its genomic information, from which can be computed the phenotypic
information of both semantic representations and patterns of stress. The
conducted experiments broadly indicated that relatively meaningful text could
be produced if the constraints on metre were relaxed, and precise metric text was
possible with loose semantic constraints, but it was difficult to produce text which
was both semantically coherent and of high quality metrically.
Keywords: natural language generation; genetic algorithms; poetry;
creative language

1. Introduction
The task of automatically generating texts that would be regarded as poetry is challenging,
as it involves many aspects of language. Moreover, there is no well-defined route by which
poems are created by human poets. However, it is possible to set down some properties
that a poetic text should manifest, which suggests that the generation might be achieved by
an algorithm that couples a suitably effective search process with a relatively separate
evaluation of the end-product; stochastic search, and in particular evolutionary
algorithms, operate in this way.
We start here with a brief survey of work on automatic poetry generation (Section 2),
before setting out a restricted definition of poetry (Section 3) as a text that embodies
meaningfulness, grammaticality and poeticness, and a model of its generation as a
stochastic search process (Section 4). We then describe McGONAGALL (Section 5), our
implemented system that adopts this model and uses genetic algorithms (Mitchell 1996) to
generate texts that are syntactically well-formed, meet certain pre-specified patterns of
metre (Attridge 1995) and broadly convey some given meaning. Finally, we present results
of some experiments conducted with McGONAGALL (Section 6) and conclude this article
with some discussion about future avenues of research (Section 7).

*Corresponding author. Email: [email protected]

ISSN 0952–813X print/ISSN 1362–3079 online
ß 2012 Taylor & Francis
http://dx.doi.org/10.1080/0952813X.2010.539029
http://www.tandfonline.com
 44 R. Manurung et al.

Wheresoever amorphous – just barely the nightclub,

howsoever apostolic amidst a calamity,
a dragon will irrigate a Copernican currant–an emphysema.
His cowlick must have incinerated a housebroken revelry as per a melamine.
your in active hydrocarbon could atone.

Figure 1. Sample output from ALFRED The Agent.

2. A brief survey of poetry generators

In this section we will examine four particular poetry generators and discuss their relative
merits, similarities, and differences. (Further reviews of this area can be found in Bailey
(1974), Gervás (2002) and Manurung (2003).)
Downloaded by [McMaster University] at 11:49 28 December 2014

2.1. ALFRED the Agent

ALFRED the Agent (Donald n.d.) is highly typical of a large number of poetry generators
that randomly choose words from a handcrafted dictionary to fill gaps in some incomplete
linguistic structure, defined either as a predefined template or via some phrase structure
rules. Figure 1 shows some ALFRED output.
Other systems of this type include Masterman’s haiku generator (Boden 1990) and
ALAMO ‘rimbaudelaires’ (Gervás 2002). Similar systems are available on the web such as
ELUAR, The Poetry Creator, and ADAM. Notable instances are RACTER and PROSE
(Hartman 1996), which have the distinction of having their poetry published – the poetry
anthology ‘The Policeman’s Beard is Half Constructed’ consisted solely of poems
constructed by RACTER. Some of these systems employ certain heuristics to simulate
the appearance of coherence and poeticness, such as assigning ad-hoc ‘emotional
categories’, e.g. {ethereality, philosophy, nature, love, dynamism} in ELUAR, and choosing
lexical items repetitively to give a false sense of coherence, as in RACTER.

2.2. Ray Kurzweil’s Cybernetic Poet

Ray Kurzweil’s Cybernetic Poet, or RKCP (Kurzweil 2001), is a commercial software
package that generates well-formed texts based on rhythmic properties of words that are
compiled in a statistical language model trained on existing poems. RKCP employs
proprietary algorithms, unavailable for us to analyse.
The RKCP text in Figure 2 is presented as a haiku, although it does not meet the
formal requirements of a haiku, i.e. lines of five, seven and five syllables. This is because
RKCP can relax constraints when it is unable to write a poem. Unlike the systems listed in
Section 2.1, RKCP attempts to conform to metre, a fundamental aspect of most poetry.

2.3. COLIBRI
COLIBRI (Diaz-Agudo, Gervás, and González-Calero 2002) uses case-based reasoning
(Aamodt and Plaza 1994) to generate formal Spanish poetry that satisfies the constraints
of strophic forms such as romances, cuartetos and tercetos encadenados. Its input is a list of
keywords representing the meaning to be conveyed, and a specification of the particular
strophic form required.
 Journal of Experimental & Theoretical Artificial Intelligence 45

Crazy moon child

Hide from your coffin
To spite your doom.

Figure 2. Sample output from Kurzweil’s Cybernetic Poet.

(a) una boca ardiente pase techo y suelo

(b) no sólo en plata o viola truncada
se vuelva mas tú y el lo juntamente
en tierra en humo en polvo en sombra en nada
(c) no sólo en boca y viola ardiente
se pase mas tú y el lo juntamente
en tierra en techo en suelo en sombra en nada
Downloaded by [McMaster University] at 11:49 28 December 2014

(d) no sólo para* boca y viola ardiente

se pase mas tú y el lo juntamente
en tia* en techo en suelo en sombra en serpiente*

Figure 3. COLIBRI: (a) input, (b) verse pattern, (c) substituted and (d) revised.

Figure 3 shows how COLIBRI works. In (a) we see keywords representing the required
meaning for the poem. In (b) a case is retrieved from a corpus of human-authored poems
with an appropriate strophic form. In (c) the keywords from (a), marked in boldface, have
been placed in the text, while maintaining syntactic well-formedness. Finally, in (d) we see
the result of a revision, where the words marked with * have been substituted in to ensure
the metre and rhyme specified by the strophic form.
COLIBRI differs from the systems in the previous sections because, as well as aiming
for syntactic well-formedness (cf. Section 2.1) and appropriate rhythm (like RKCP), its
goal is a text that is ‘about’ a given message, even if this is approximated rather trivially.
However, when COLIBRI revises a poem to maintain metre and rhyme it pays no heed to
meaning. Modifications to the text at this stage may destroy whatever meaning has already
been set up (cf. Section 4.1).

2.4. Chart generation of rhythm-patterned text

Our initial attempt at poetry generation (Manurung 1999) constructs a syntactically
well-formed text that conveys a conjunction of atomic formulae and satisfies a given
metre. A context-free grammar guides exhaustive search for phrases that match the
semantic items, with chart generation (Kay 1996) used to reduce the computational
complexity. Patterns of required rhythms act as filters for potential structures during
generation. Figure 4 shows a generated text in (b) with the metre of a limerick and the
semantics in (a).
Like COLIBRI, this system takes account of syntax, rhythm and meaning, but its
account of meaning is much more sophisticated. In particular, it conveys the predicate-
argument structure of the input. Also, COLIBRI handles meaning and metre separately,
but our chart generator maintains syntactic, semantic and rhythmic well-formedness at
every step. Unfortunately, due to its exhaustive search, it is computationally very
 46 R. Manurung et al.

(a) cat(c) ∧ dead(c) ∧ eat(e,c,b) ∧ past(e) ∧ bread(b) ∧ gone(b)

(b) The cat is the cat which is dead.
The bread which is gone is the bread.
The cat which consumed
the bread is the cat
which gobbled the bread which is gone.

Figure 4. Sample (a) input semantics and (b) output limerick from our chart generator.

expensive, even in our tests on a tiny ‘toy’ grammar and lexicon. It is also not very flexible,
as it can generate only immaculate results – if the linguistic resources are insufficient to
produce the required output, it fails completely. We believe that a poetry generator should
be robust, in the sense that it produces the best possible output with the resources at its
Downloaded by [McMaster University] at 11:49 28 December 2014

disposal.

3. Defining poetry
There is no universally accepted definition of what counts as poetry, but we will argue for
the use of a particular restricted definition in our investigations.

3.1. The unity of poetry

‘Poetry is a literary form in which language is used in a concentrated blend of sound and
imagery to create an emotional response.’ (Levin 1962)
This points to a strong interaction, or unity, between the form and content of a poem.
The diction and grammatical construction of a poem affects the message that it conveys to
the reader over and above its obvious denotative meaning. This contrasts with the
commonly-held linguistic notion that a text serves as a medium for conveying its semantic
content.

3.2. Abusing poetic license

Boden claims, in Boden (1990), that:
‘Readers of [poetry] are prepared to do considerable interpretative work. . . . In general, the
more the audience is prepared to contribute in responding to a work of art, the more chance
there is that a computer’s performance (or, for that matter, a human artist’s) may be
acknowledged as aesthetically valuable . . . Much of the beauty, might one say, is in the eye of
the beholder . . . Hence poetry, especially poetry of a highly minimalist type (such as haiku),
is more tolerant of programmed production than prose is.’
The implication here is that the automatic generation of poetry is relatively easy
because the burden of assigning meaning to a text is carried by the reader. We believe that
this argument abuses poetic license1, and that by taking it to the extreme, one can
potentially justify any randomly created text as being poetry. This would seriously
undermine any research on poetry generation. When appraising the output of a computer
 Journal of Experimental & Theoretical Artificial Intelligence 47

The Lion, the Lion, he dwel ls in the waste,

He has a big head and a very smal l waist;
But his shoulders are stark, and his jaws they are grim,
And a good little child wil l not play with him.

Figure 5. Hillaire Belloc’s ‘The Lion’.

program that is claimed to generate poetry, to what extent can we allow deviations
(e.g. semantic incoherence) in the text, by invoking ‘poetic license’?
We believe the opposite of Boden’s claim to be the case: automatic generation of poetry
is harder than that of prose. Poetry is further constrained by rules of form that prose need
not adhere to. Using true poetic license to turn a phrase in an effective way requires even
greater mastery of the language than that for producing prose. Deviations from the rules
Downloaded by [McMaster University] at 11:49 28 December 2014

and norms must have some purpose, and not be random.

3.3. A restricted definition of poetry

We must have a definition of poetry that renders falsifiable the claim that a program
successfully generates poetry. Our proposed definition is: a poem is a natural language
artifact that simultaneously satisfies the constraints of grammaticality, meaningfulness
and poeticness.
. Grammaticality: A poem must be syntactically well-formed. This might seem
obvious for natural language texts, but it must be stated explicitly in the context
of poetry. Syntactic well-formedness in poetry may well be different from that of
ordinary texts (e.g. figurative language), but it is still governed by rules.
. Meaningfulness: A poem must intentionally convey some conceptual message that
is meaningful under some interpretation. By stipulating this constraint, we are
more critical of what we consider not to be poetry, e.g. the output of ALFRED
(Figure 1), RKCP (Figure 2) and, to a lesser extent, COLIBRI (Figure 3).
. Poeticness: A poem must exhibit features that distinguishes it from non-poetic
text. In our research, we concentrate on concretely observable aspects, specifically
the rhythmic patterns, or metre, of a poem.
This characterisation suggests a ‘classical’ account of poetry, i.e. one where adherence
to regular patterns in form is paramount. This avoids some of the complications of
imagery or interpretation that are central to assessing more free forms of verse, and
ensures that many of the important aspects of a text can be defined formally. It is an
approximate formalisation of the sort of verse that was normal within British culture until
the mid-20th century and is still commonplace today, and includes established forms such
as limericks and sonnets (although we have not yet tackled the important issue of rhyme).
Figure 5 exhibits Hillaire Belloc’s ‘The Lion’, an example of this genre – and one that we
will revisit throughout this article.
Although our definition does impose constraints, it says nothing about abstract
notions such as imagery and metaphor, which are beyond the scope of our current work.
As the field advances, these might become further requirements, or be incorporated under
the poeticness constraint.
Table 1 shows how the generators in Section 2 fit our restricted definition.
 48 R. Manurung et al.

Table 1. Poetry generators (Section 2) and how they address our constraints.

System Grammaticality Poeticness Meaningfulness

ALFRED Yes No No
RKCP Yes Yes No
COLIBRI Yes Yes Yes
Chart generation Yes Yes Yes

4. Evolving poetry
We now turn to the question of how to generate texts automatically that satisfy the three
constraints of grammaticality, meaningfulness and poeticness.
Downloaded by [McMaster University] at 11:49 28 December 2014

4.1. The problem with unity

Poetry generation can be viewed as a specialised instance of natural language generation, or
NLG, a subfield of artificial intelligence that focuses on the development of computer
systems that can produce understandable texts in a human language, starting from some
nonlinguistic representation of a communicative goal as input (Reiter and Dale 2000).
A considerable amount of research has been done in NLG on the so-called ‘generation
gap’ problem, where interdependent decisions must be made across various levels of
linguistic representation (Meteer 1991). This problem is exacerbated by the unity of poetry
(Section 3.1). For example, with regards to metre, every single linguistic decision
potentially determines the success of a poem in fitting a regular defined pattern. Various
solutions have been proposed, such as: using alternative system architectures, using better
search methods and employing a new paradigm of overgeneration. We have drawn upon
this research to propose our model of poetry generation as stochastic search, which we
present in the following section.

4.2. Poetry generation as stochastic search

Poetry generation can be viewed as a state space search problem, where a state in the
search space is a possible text with all its underlying representation, from semantics all the
way down to phonetics. A goal state satisfies the three constraints of meaningfulness,
grammaticality and poeticness. Such a search space is undoubtedly immense, even – given
the recursive structures in natural language – infinite. Our proposed solution is to employ
a form of stochastic search, a heuristic search strategy that relies on random traversal of a
search space with a bias towards more promising solutions. It has become increasingly
popular for solving computationally hard combinatorial problems from AI domains such
as planning, scheduling and constraint satisfaction, and has been shown to outperform
deterministic approaches in a number of domains.
The particular search methodology that we use is the genetic algorithm, which is
essentially an iteration of two phases, evaluation and evolution, applied to an ordered set,
called the population, of candidate solutions (Mitchell 1996).
 Journal of Experimental & Theoretical Artificial Intelligence 49

4.3. Satisfying grammaticality, meaningfulness, and poeticness

How do we design our genetic algorithm to find solutions that satisfy the three constraints
described in Section 3.3? There are two ways that constraints can be implemented: either
invalid solutions are totally excluded from the search space, or they are admitted to the
space but with a bias against them (Davis and Steenstrup 1987). Our three constraints are
not independent. Within a generator, we must be able to relate syntactic structure and
semantic content systematically, probably by some variant of the widely accepted notion
of compositional semantics. This would be hard to arrange if we allowed syntactically
ill-formed structures. Also, the kind of poeticness that we are concentrating on (metre) can
be imposed regardless of grammaticality, but (as we shall see when examining The Lion)
grammaticality may sometimes take precedence. We have therefore decided to treat
grammaticality as a prerequisite for meaningfulness and poeticness; that is, we will
implement grammaticality as an absolute constraint on the items admitted to the search
Downloaded by [McMaster University] at 11:49 28 December 2014

space, through the design of representation and genetic operators. In doing this, we can
call on the immense amount of linguistic work on mechanisms for ensuring
grammaticality.
The remaining constraints of meaningfulness and poeticness will be implemented as
preferences via the evaluation functions of the genetic algorithm.

5. McGONAGALL, an instance of an evolutionary poet

McGONAGALL is our implementation of poetry generation as stochastic search. In this
section we will describe how McGONAGALL ensures grammaticality through its
representation (Section 5.1) and genetic operators (Section 5.2), and how it optimises
poeticness (Sectction 5.3) and meaningfulness (Section 5.4).

5.1. Linguistic Representation

Linguistic structures are represented in McGONAGALL using lexicalised tree-adjoining
grammar, or LTAG (Schabes 1990), augmented with the use of feature structures
(Vijay-Shanker and Joshi 1988). These grammars are based on the composition of
elementary trees, of which there are two types, i.e. initial trees and auxiliary trees. These
trees represent minimal linguistic structures in the sense that they account for all and only
the arguments of the head of the syntactic constituent they represent, e.g. a sentential
structure would contain a verb and all its complements. There are two types of operations,
substitution of initial trees (typically labelled ) and adjunction of auxiliary trees (typically
labelled ), that can be used to compose larger tree structures.
In LTAG, the derivation tree is a data structure that records the composition of
elementary trees using substitution and adjunction. The derived tree, on the other hand, is
the phrase structure tree created by performing all the operations specified within the
deriviation tree; in a sense, the derived tree is the result of generation, whereas
the derivation tree is a trace of how it was built up. Figure 6 gives an example of this.
The derived tree in (c) is obtained by a series of operations on the elementary trees in (a).
These operations are recorded by the derivation tree in (b). Each node in the derivation
tree specifies an elementary tree. Furthermore, each node aside from the root node also
specifies the Gorn address2 of the node in its parent’s elementary tree where the operation
occurs, and is indicated by the label of the edge connecting a node to its parent.
 50 R. Manurung et al.

aread aJohn b yesterday abook aa

b yesterday
S
S
NP S NP aread
D S
NP VP
Adv S*
N Adv S* D N

a NP VP
V NP
Yesterday
John Yesterday Book

V NP
Read
(a)
S
aJohn Read
NP

Adv S
abook
Downloaded by [McMaster University] at 11:49 28 December 2014

N
aread NP

Yesterday NP VP
1 2.2 0 John
D N
aJohn abook b yesterday N V NP aa
1 D Book

John Read D N
aa
a

a Book
(b) (c) (d)

Figure 6. A derivation tree captures the process of composition of elementary trees.

For example, book subtitutes into read at address 2.2, the object NP node. The process of
composing the elementary trees is illustrated in (d).
The derivation tree can therefore be seen as the basic formal object that is constructed
during sentence generation from a semantic representation (Joshi 1987), and is the
appropriate data structure on which to perform nonmonotonic operations in our
stochastic generation framework (Section 5.2). Essentially, the LTAG derivation tree
forms the genotypic representation of our candidate solutions, from which we can compute
the phenotypic information of semantic (Section 5.4) and prosodic (Section 5.3) features via
the derived tree.
We adopt a simple ‘flat’ semantic representation (Hobbs 1985) that is often used in
NLG (Koller and Striegnitz 2002). A semantic expression is a set of first order logic literals,
which is logically interpreted as a conjunction of all its members. The arguments of these
literals represent concepts in the domain such as objects and events, while the functors
state relations between these concepts. For example, the representation of the semantics of
the sentence ‘John loves Mary’ is { john( j), mary(m), love(l, j, m)}, where l is the event of j,
who has the property of ‘being John’, loving m, who has the property of ‘being Mary’.
The semantic expression of a tree is the union of the semantic expressions of its
constituent elementary trees, with appropriate binding of variables during substitution and
adjunction through the use of semantic signatures to control predicate-argument structure;
cf. Stone and Doran (1997) and Kallmeyer (2002). Ultimately semantic information
 Journal of Experimental & Theoretical Artificial Intelligence 51

αloves ; S αloves ={loves(X,Y,Z)} S ={loves(X,Y,Z), john(_,Y), mary(,Z)}

S X S X

NP Y VP X NP Y VP X

V X NP Z N Y V X NP Z

{A/Y} Loves X,Y,Z John _,Y Loves X,Y,Z N Z

{B/Z}

Mary _,Z
α John ; SαJohn ={john(_,A)} αMary; SαMary ={mary(_,B)}

NP A NP B
Downloaded by [McMaster University] at 11:49 28 December 2014

N A N B

John _,A Mary _,B

Figure 7. Compositional semantics through unification of signatures.

originates from the lexicon, where words are annotated with the meanings they convey.
Figure 7 shows an example of the substitution of two noun phrases as the subject and
object of a transitive tree. During the substitution of John at the subject NP position, the
signatures Y and A are unified. Likewise, for the substitution of Mary at the object NP
position, the signatures Z and B are unified. From this, we can recover the semantics of
the derivation S ¼ {love(X, Y, Z), john(_, Y), mary(_, Z)}.
Each word is associated with its phonetic spelling, taken from the CMU pronouncing
dictionary (Weide 1996). Vowels are marked for lexical stress, with 0 for no stress, 1 for
primary stress and 2 for secondary stress. For example, the spelling of ‘dictionary’ is
[D, IH1, K, SH, AH0, N, EH2, R, IY0]. For monosyllables, closed class words (e.g. the)
receive no stress, and open class words (e.g. cat) receive primary stress.

5.2. Genetic operators for linguistic structure building

We have implemented baseline (‘blind’) mutation operators that randomly add or delete
subtrees of a derivation tree. In addition, we have also implemented a subtree swapping
operator, which randomly swaps two derivation tree subtrees. When these subtrees belong
to the same derivation tree, it behaves as a mutator, otherwise, it is used as a crossover
operator. These nonmonotonic tree-based operators are similar to those commonly found
in genetic programming (Angeline 1996).
We also implemented semantically motivated (‘smart’) genetic operators that exploit
domain knowledge and attempt to maximise meaningfulness by conveying the target
semantics (Section 5.4), while still maintaining grammaticality. This is motivated by NLG
systems such as PROTECTOR (Nicolov, Mellish, and Ritchie 1995) and SPUD (Stone and
Doran 1997). When adding content, these operators attempt to convey the unrealised
semantics, i.e. the portion of the target semantics that has not yet been realised by the
 52 R. Manurung et al.

The Lion, the Lion, he dwells in the waste,

He has a big head and a very small waist;
But his shoulders are stark, and his jaws they are grim,
And a good little child • will not play with him.

Figure 8. Metre pattern of Hillaire Belloc’s ‘The Lion’.

(a) (b) (c)

[w,s,w,w,s,w,w,s,b, [x,x,x,x,x,b, [w,s,w,w,s,w,w,s,w,w,s,b,
w,s,w,w,s,w,w,s,b, x,x,x,x,x,x,x,b, w,s,w,w,s,w,w,s,w,w,s,b,
w,s,w,w,s,b, x,x,x,x,x,b] w,s,w,w,s,w,w,s,w,w,s,b,
w,s,w,w,s,b, w,s,w,w,s,w,w,s,w,w,s,b]
w,s,w,w,s,w,w,s,b]
Downloaded by [McMaster University] at 11:49 28 December 2014

Figure 9. Target forms for (a) a limerick, (b) a haiku, and (c) ‘The Lion’.

candidate solution. Conversely, when deleting content, operators try to remove extraneous
items, i.e. semantics realised by the candidate solution but not present in the target
semantics.

5.3. Evaluating poeticness

In McGONAGALL, poeticness is taken to be the (well-defined and objectively observable)
metre – regular patterns in the rhythm of the lines (Attridge 1995). Figure 8 shows the
metre of Belloc’s ‘The Lion’, with stressed syllables in bold type, unstressed syllables in
normal type, syllables extraneous to the underlying metre in italics, and . indicating a
‘missing’ syllable. Each line has the same number of stressed syllables, with a regular
pattern of ‘beats’ at intervals of two unstressed syllables.
McGONAGALL tries to maximise the similarity between the target form, a specification
of the required metrical constraints, and the candidate form, the stress pattern of a
candidate solution.
The target form is encoded as a list of target syllables, notated as follows: w (‘weak’) is
an unstressed syllable, s (‘strong’) is a stressed syllable, x (‘wildcard’) is any syllable, and b
indicates a linebreak. Figure 9 shows example target forms for a limerick, a haiku and
‘The Lion’ (formatted into lines for the purpose of readability).
The candidate form, a representation of the metre exhibited by a candidate solution, is
encoded as a list of candidate syllables obtained from a derived tree by concatenating the
phonetic spellings at its leaf nodes (Section 5.1). The notation is as follows: 0 is an
unstressed syllable, 1 a primary stressed syllable, 2 a secondary stressed syllable and b is a
potential site for a natural line break. In addition, candidate syllables 0, 1 and 2 have a
subscript index of either 1 or n, indicating whether they belong to a monosyllabic or
polysyllabic word, respectively. Figure 10 shows the derived tree representing the text
‘The lion has a big head’, and how its stress pattern is obtained.
Having explicit linebreaks in both target and candidate allows for a flexible handling of
linebreaking anywhere in the text, in contrast to a hierarchical model of metrical forms
consisting of lines of syllables (e.g. in Gervás 2000; Diaz-Agudo et al. 2002).
To compute the similarity between the target and candidate forms, we use the minimum
edit distance (Sankoff and Kruskal 1983), in which the distance between two strings is the
 Journal of Experimental & Theoretical Artificial Intelligence 53

Derived tree: Lexicon:

Poem
Ortho Pronunciation Lexical stress
the [DH,AH0] 01
S SB lion [L,AY1,AH0,N] 1 n ,0 n
has [HH,AH1,Z] 11
a [AH0] 01
NP VP . big [B,IH1,G] 11
head [HH,EH1,D] 11
. [-] b
D N V NP

The Lion has D NP

a Adj NP
Downloaded by [McMaster University] at 11:49 28 December 2014

Candidate form:

Big N
[ 0 1 , 1 n , 0 n , 1 1 , 0 1, 1 1 , 1 1 , b ]

Head

Figure 10. The candidate form of a text is a concatenation of lexical stress.

minimal sum of costs of operations (symbol insertion, deletion and substitution) that
transform one string into another. The minimum edit distance can be efficiently computed
(Jurafsky and Martin 2000) in a way that produces a pairwise syllable alignment between
candidate and target, thus indicating the operations that yield the minimum cost. The costs
are shown in Table 2(a)–(c), which reflect our intuitions, as follows:
(i) There should be no penalty for aligning candidate unstressed, stressed or
linebreak syllables with target syllables of that same class, nor where
non-linebreak candidate syllables align with a wildcard in the target.
(ii) Linebreaks cannot align with non-linebreaks, in either direction.
(iii) Having a linebreak in the candidate where none exists in the target (enjambment,
the running on of a sentence from one line of a poem to the next) should be
relatively expensive. Allocating no penalty for deletion of a candidate linebreak
lets a line consist of more than one clause.
(iv) All other operations should incur some penalty. Insertion and deletion should be
costlier for stressed syllables than for unstressed syllables.
Our candidate forms indicate lexical stress patterns as if the words were pronounced in
isolation. Within poetic text, context can affect stress. To compensate for this, the system
iterates over the minimum edit distance alignment, detecting certain patterns and adjusting
the similarity value. Two types of patterns are implemented: two consecutive deletions, or
two consecutive insertions, of non-linebreaks, increases the cost by 1; the destressing of a
stressed candidate syllable adjacent to a stressed target syllable, or the stressing of an
unstressed candidate syllable adjacent to an unstressed target syllable, decreases the
cost by 1.
Our metre evaluation function, F metre, takes the value computed by the minimum edit
distance algorithm, adjusts it using our context-sensitive compensation scheme and
normalises it to the interval [0, 1].
 54 R. Manurung et al.

Table 2. Operation costs for (a) substitution, (b) insertion, and (c) deletion.

(a)

Cost w s x b

01 0 2 0 1
0n 0 2 0 1
11 3 0 0 1
1n 3 0 0 1
2n 1 1 0 1
b 1 1 1 0
(b)

Cost
Downloaded by [McMaster University] at 11:49 28 December 2014

w 1
s 3
x 1
b 10
(c)

Cost
01 1
0n 1
11 3
1n 3
2n 2
b 0

Table 3. Metre similarity fitness scores for various candidates.

Candidate text Candidate form F metre

The Lion, the Lion, he dwells in the [01, 1n, 0n, b, 01, 1n, 0n, b, 01, 11, 01, 01, 0.787
waste. He has a big head and a very 11, b, 01, 11, 01, 11, 11, 01, 01, 1n, 0n,
small waist. But his shoulders are 11, 11, b, 01, 01, 1n, 0n, 01, 11, b, 01,
stark, and his jaws they are grim, 01, 11, 01, 01, 11, b, 01, 01, 11, 1n, 0n,
and a good little child will not play 11, 01, 01, 11, 01, 01, b]
with him.
There was an old man with a beard, who [01, 01, 01, 11, 11, 01, 01, 11, b, 01, 11, b, 0.686
said, ‘it is just as i feared! two owls 01, 01, 11, 01, 01, 11, b, 11, 11, 01, 01,
and a hen, four larks and a wren, have 11, b, 11, 11, 01, 01, 11, b, 11, 11, 11,
all built their nests in my beard!’ 01, 11, 01, 01, 11, b]
Poetry is a unique artifact of the human [1n, 0n, 0n, 01, 01, 0n, 1n, 1n, 0n, 2n, 01, 0.539
language faculty, with its defining 01, 1n, 0n, 1n, 0n, 1n, 0n, 0n, b, 01, 01,
feature being a strong unity between 0n, 1n, 0n, 1n, 0n, 1n, 0n, 01, 11, 1n,
content and form. 0n, 0n, 0n, 1n, 1n, 0n, 01, 11, b]
John loves Mary. [11, 11, 1n, 0n, b] 0.264
 Journal of Experimental & Theoretical Artificial Intelligence 55

Table 4. Semantic similarity fitness scores for various candidates.

Candidate text Candidate semantics F sem

The lion has a big head. {lion1(_, L), own(_, L, H ), 0.525

head(_, H ), big(_, H )}
The lion has a head and a waist {lion(_, L), own(_, L, H ), head(_, H ), 0.598
own(_, L, W ), waist(_, W )}
The lion dwells in the waste {lion(_, L), dwell(D, L), 0.078
inside(_, D, W ), waste(_, W )}
John loves Mary { john(_, J ), love(_, J, M), mary(_, M)} 0.0451
John and Mary love the lion’s big head { john(_, J ), love(_, J, H ), mary(_, M), 0.389
love(_, M, H ), lion(_, L), own(_, L, H ),
head(_, H ), big(_, H )}
Downloaded by [McMaster University] at 11:49 28 December 2014

Table 3 shows F metre values for various candidate texts, against the target form in
Figure 9(c). The first is Belloc’s actual poem, which itself contains some metrical
imperfections; the second is a limerick by Edward Lear; the third is an extract from an
academic text, containing roughly the correct number of syllables; the last is chosen for its
inappropriateness. The F metre scores do not conflict radically with our intuitions of poetic
metre.

5.4. Evaluating meaningfulness

The McGONAGALL approach to meaningfulness is similar to the above approach to
poeticness: try to maximise the similarity between the target semantics, a specification of
the meaning an optimal solution should convey, and a candidate semantics, the meaning
conveyed by a candidate solution. This requires a method for computing the similarity
between two semantic expressions.
Following Love (2000), we propose two factors that must be considered: structural
similarity and conceptual similarity. Structural similarity measures the degree of
isomorphism between two semantic expressions. Conceptual similarity is a measure of
relatedness between two concepts (logical literals). We simply use the following: two
concepts are the same if and only if they share the same literal functor.
Computing a structural similarity mapping between two expressions is an instance of
the NP-hard graph isomorphism problem. However, we have implemented a greedy
algorithm that serves our purposes and runs in O(N3), based on Gentner’s structure
mapping theory (Falkenhainer, Jorbus, and Gentner 1989). It takes two sets of logical
literals, Starget and Scandidate, and attempts to ‘align’ the literals. In doing this, it creates
various variable bindings and also two sets of ‘dangling’ literals that are left over (from
Starget and Scandidate) as unmatched items.
We then apply a function F sem, normalised to [0, 1], to compute a score based on
various aspects of the best match that has been achieved; this is based on Love’s
computational model of similarity (Love 2000).
Table 4 shows an example of computing semantic similarity for a selection of candidate
texts against a target semantic that represents the second line of Belloc’s ‘The Lion’, i.e. ‘[The
lion] has a big head and a very small waist’. The target semantic expression is as follows:
Starget ¼ flionð , LÞ, ownð , L, H Þ, head ð , H Þ, bigð , H Þ,
ownð , L, W Þ, waistð , W Þ, small ðS, W Þ, veryð , S Þg
 56 R. Manurung et al.

The first two texts convey a subset of the target; the third text conveys an altogether
different fact about the lion; the fourth text is purposely inappropriate; and the last text,
interestingly, conveys the semantics of the first text in its object to the verb ‘love’. As with
our metre similarity function, we believe that the F sem scores roughly approximate human
intuitions.

6. Experiments with McGONAGALL

We conducted five tests, which clustered into three stages of experiments:
. Metrical generation. We set the program to generate grammatical texts that satisfy
a given target metre. This measures its ability to generate solutions that optimise
poeticness without concern for meaning. The evaluation function used was F metre
(Section 5.3). The target form used was that of a limerick, as in Figure 9(a).
Downloaded by [McMaster University] at 11:49 28 December 2014

. NLG test. This test measured the ability of McGONAGALL to generate solutions
that optimise meaningfulness, without concern for metre. The evaluation function
was F sem (Section 5.4). This effectively causes McGONAGALL to behave as a
conventional NLG system, i.e. it attempts to produce a text that conveys a given
input semantics. The target semantics used is a representation of the first two lines
of ‘The Lion’, with a slight alteration where we have replaced the original opening
noun phrase ‘the lion, the lion’ with ‘the african lion’, i.e. ‘The african lion, he dwells
in the waste. He has a big head and a very small waist’. The target semantic
expression is as follows:
Starget ¼ flionð , l Þ, africanð , l Þ, dwell ðd, l Þ, insideð , d, wasÞ, wasteð , wasÞ,
ownð , l, hÞ, head ð , hÞ, bigð , hÞ, ownð , l, waiÞ, waistð , waiÞ,
small ðs, waiÞ, veryð , sÞg

We conducted two variants of this test, one with the syntactically-aware ‘blind’
operators and one with the syntactically and semantically-aware ‘smart’ operators
(Section 5.2).
. Poetry generation. In this test we measured the ability of McGONAGALL to perform
the task it was designed for, namely the generation of texts that simultaneously
satisfy grammaticality, meaningfulness and poeticness. We took a very simple
approach to combining the metre similarity and semantic similarity functions –
the arithmetic mean of their scores, i.e. F poetry ¼ (F metre þ F sem)/2. We used the
same target form and target semantics as in the previous tests. As in the previous
test, two variants were conducted: with the ‘blind’ operators and with the ‘smart’
operators.

6.1. GA parameters
Throughout our testing, we employed one of the simplest selection algorithms,
proportionate selection, which assigns a distribution that accords parents a probability
to reproduce that is proportional to its fitness (Bäck, Fogel, and Michalewicz 1997).
Individuals are sampled from this distribution using stochastic universal sampling, which
minimises chance fluctuations in sampling (Baker 1987). To reduce the chances of
premature convergence or stagnation, we used an elitist population of 20% of the entire
 Journal of Experimental & Theoretical Artificial Intelligence 57

Table 5. Statistical summary of test results.

Test Min Max Mean SD

Stage 1:
Metrical generation 0.86 1.00 0.93 0.05
Stage 2:
NLG test (blind operators) 0.74 0.93 0.85 0.06
NLG test (smart operators) 0.93 0.99 0.97 0.03
Stage 3:
Poetry generation test (blind operators) 0.60 0.81 0.69 0.06
Poetry generation test (smart operators) 0.75 0.83 0.79 0.02
Downloaded by [McMaster University] at 11:49 28 December 2014

population, which was found to yield the best results during an initial experiment with
different ratios. The population size was set to 40, inspired by Cheng (2002), who in turn
points to empirical studies by Goldberg (1989). Each test was run five times, and each run
lasted for 500 iterations.
The three mutation operators used, along with their probabilities, were creation (0.5),
adjunction (0.3), and deletion (0.2). The choice of these probabilities is intended to account
for the fact that the substitution operation typically introduces the essential linguistic
structures, and that adjunction is for auxiliary constituents. For crossover, the subtree
swapping operator was used. The probabilities of applying genetic operators were
pmutation ¼ 0.6, pcrossover ¼ 0.4, for both the blind and smart variants.
A handcrafted grammar and lexicon was used, with 33 elementary trees and 134 lexical
items, 28 of which were closed class words. Most of the content words were taken from
‘The Bad Child’s Book of Beasts’ (Belloc 1991).

6.2. Overview of results and analysis

Table 5 shows a statistical summary of the best fitness scores obtained during the three
stages of experiments. It shows the minimum, maximum, mean and standard deviation of
the best fitness scores obtained after the last iteration across the five runs of each test.
From this table we can observe that the GA is able to find optimal solutions with fitness
approaching the maximum of 1.0 during the first two stages of experiments, but is failing
to do so for the difficult multi-objective optimisation task of the third stage.
Figure 11 shows, over time, the maximum and average of the best fitness scores across
all the GA runs for the metrical generation test, (a), NLG test with blind and smart
operators, (b) and (c), and poetry generation test with blind and smart operators,
(d) and (e).

6.3. Metrical generation

Table 6 shows the highest-scoring candidate from the metrical generation test.
As expected, the text is syntactically well-formed, being generated through a composition
of LTAG elementary trees. The target metre is satisfied perfectly in terms of syllable count,
stress placement and linebreaking.
 58 R. Manurung et al.

(a)
1

0.8

Fitness score
0.6

0.4

0.2
Maximum
Average
0
0 50 100 150 200 250 300 350 400 450 500
Iteration
(b) (c)
1 1

0.8 0.8
Downloaded by [McMaster University] at 11:49 28 December 2014

Fitness score

Fitness score
0.6 0.6

0.4 0.4

0.2 0.2
Maximum Maximum
Average Average
0 0
0 50 100 150 200 250 300 350 400 450 500 0 50 100 150 200 250 300 350 400 450 500
Iteration Iteration
(d) (e)
1 1
Fitness score

0.8 0.8
Fitness score

0.6 0.6

0.4 0.4

0.2 0.2
Maximum Maximum
Average Average
0 0
0 50 100 150 200 250 300 350 400 450 500 0 50 100 150 200 250 300 350 400 450 500
Iteration Iteration

Figure 11. Maximum and average of best fitness scores for (a) metrical generation test, NLG test
with (b) blind and (c) smart operators, poetry generation test with (d) blind and (e) smart operators.

Table 6. Best found solution for metrical generation test.

Fitness score 1.00

Text They play. An expense is a waist.

A lion, he dwells in a dish.
He dwells in a skin.
A sensitive child,
he dwells in a child with a fish.
 Journal of Experimental & Theoretical Artificial Intelligence 59

Table 7. Best found solution for NLG test with blind operators.

Fitness score 0.93

Text An african lion, it dwells in a waste.

Its big head, that is him, it has its very small waist.
Matched {(inside(_2, d, was), inside(_47, _45, _48)), (lion_0, l), lion(_50, _46)),
literals (african(_1, l), african(_51, _46)), (dwell(d, l), dwell(_45, _46)),
(waste(_3, was), waste(_49, _48)), (head(_5, h), head(_59, _53)),
(own(_4, l, h), own(_61, _46, _53)), (own(_7, l, wai), own(_56, _46, _54)),
(big(_6, h), big(_60, _53)), (small(s, wai), small(_57, _54)),
(waist(_8, wai), waist(_55, _54)), (very(_9, s), very(_58, _57))}
Unmatched ;
target
Unmatched {own(_52, _53, _54)}
candidate
Downloaded by [McMaster University] at 11:49 28 December 2014

6.4. NLG test with blind operators

Table 7 shows the highest-scoring candidate from the blind operators NLG test. It shows
the fitness score, text and the best mapping of Starget to Scandidate in terms of matched and
unmatched literals.
The generator achieves a near-optimal solution of 0.93. At a glance, we can note that the
text is certainly more ‘about’ the concepts within the target semantics than the best solution
from the metrical generation test (Table 6) – at the expense of metre. This shows the marked
effect of changing the evaluation function from metre similarity to semantic similarity.
In terms of the evaluated semantic mapping, the only suboptimal feature is the presence of
the unmatched candidate literal own(_52, _53, _54), which represents the fact that the head
owns the waist. This is conveyed by the peculiar second sentence. If we observe the text itself,
we can see that the facts that the head and waist belong to the lion are expressed by the
genitive determiner ‘its’. There is also the spurious relative clause ‘that is him’ that does not
contribute to the semantics, but makes the text unnecessarily complex.

6.5. NLG test with smart operators

Table 8 shows the highest-scoring candidate from the smart operator NLG test. It shows
that an optimal solution was found; i.e. a text that conveys the semantics without any
unmatched literals (the score of 0.99 and not 1 is caused by rounding off in the scaling
function). Our own subjective analysis of the text itself seems to confirm this: the text is
clearly understandable, if somewhat verbose.
Comparing Figure 11(b) to (c) we can see that the semantically smart operators lead to
these optimal solutions much more rapidly and consistently.
Although McGONAGALL successfully found an optimal solution in this instance, some
further tests indicated that it struggles to do so when the size of Starget is roughly double
the size of the one used in this test.

6.6. Poetry generation with blind operators

Table 9 shows the highest-scoring candidate from the blind operator poetry generation
test. It shows the fitness score, text and semantic mapping of Starget to Scandidate. In contrast
 60 R. Manurung et al.

Table 8. Best found solution for NLG test with smart operators.

Fitness score 0.99

Text An african lion, who has a very small waist,

it dwells in a waste. Its head, it is big.
Matched {(inside(_2, d, was), inside(_45, _43, _46)), (lion(_0, l), lion(_48, _44)),
literals (african(_1, l), african(_49, _44)), (dwell(d, l), dwell(_43, _44)),
(waste(_3, was), waste(_47, _46)), (head(_5, h), head(_57, _56)),
(own(_4, l, h), own(_58, _44, _56)), (own(_7, l, wai), own(_50, _44, _51)),
(big(_6, h), big(_55, _56)), (small(s, wai), small(_53, _51)),
(waist(_8, wai), waist(_52, _51)), (very(_9, s), very(_54, _53))}
Unmatched ;
target
Unmatched ;
candidate
Downloaded by [McMaster University] at 11:49 28 December 2014

Table 9. Best found solution for poetry generation test with blind operators.

Fitness score 0.81

Text A lion, it dwells in a waste.

A lion, it dwells in a waste.
A waste will be rare.
Its head will be rare.
Its waist, that is small, will be rare.
Matched {(head(_5, h), head(_53, _52)), (own(_4, l, h), own(_54, _28, _52)),
literals (own(_7, l, wai), own(_48, _28, _45)), (lion(_0, l), lion(_29, _28)),
(dwell(d, l), dwell(_27, _28)), (inside(_2, d, was), inside(_30, _27, _31)),
(waste(_3, was), waste(_32, _31)), (small(s, wai), small(_47,_45)),
(waist(_8, wai), waist(_46, _45))}
Unmatched {african(_1, l), big(_6, h), very(_9, s)}
target
Unmatched {rare(_33, _34), will(_35, _36), waste(_37, _34), dwell(_38, _39),
candidate lion(_40, _39), inside(_41, _38, _42), waste(_43, _42), rare(_44, _45),
will(_49, _50), rare(_51, _52), will(_55, _56)}

to the previous tests, no optimal solution was found to this much more difficult
multi-objective optimisation task.
The text is metrically perfect. However, in terms of conveying Starget, it performs
suboptimally. The unmatched target literals show that the text is failing to convey three
concepts from Starget, i.e. that the lion is african, that its head is big, and that the waist is
very small. The text contains three instances of the phrase ‘will be rare’, which is unrelated
to the semantics of Starget. While this phrase contributes positively to achieving the target
metre, it leads to unmatched candidate literals. Also, the second line is a repeat of the first.
Although it contributes positively to the metre, our mapping algorithm is unable to match
the semantics it conveys, resulting in unmatched candidate literals that penalises the
semantic similarity score. This does not seem to reflect the intuition that most humans
would have reading the text, i.e. that it is simply a repetition of the first line.
 Journal of Experimental & Theoretical Artificial Intelligence 61

Table 10. Best found solution for poetry generation test with smart operators.

Fitness score 0.83

Text A very . african lion,

who is african, dwells in a waste.
Its head, that is big,
is very . big.
A waist, that is its waist, it is small.
Matched {(inside(_2, d, was), inside(_150, _147, _149)), (lion(_0, l), lion(_151, _140)),
literals (dwell(d, l), dwell(_147, _140)), (waste(_3, was), waste(_148, _149)),
(head(_5, h), head(_138, _136)), (own(_4, l, h), own(_139, _140, _136)),
(own(_7, l, wai), own(_146, _140, _143)), (small(s, wai), small(_142, _143)),
(african(_1, l), african(_152, _140)), (big(_6, h), big(_135, _136)),
(waist(_8, wai), waist(_144, _143))}
Unmatched {very(_9, s)}
Downloaded by [McMaster University] at 11:49 28 December 2014

target
Unmatched {very(_137, _135), big(_141, _136), waist(_145, _143), very(_153, _152),
candidate african(_154, _140)}

The original text that conveys Starget has 22 syllables, whereas a limerick has
34 syllables. This may explain why the generator had to ‘pad out’ the remaining syllables
by adding unrelated content such as ‘will be rare’ and repeating the first line, to the
detriment of semantic similarity.

6.7. Poetry generation with smart operators

Table 10 shows the highest-scoring candidate from the smart operators poetry generation
test. Again, no optimal solution is found. However, although the fitness score is very
similar to the one in Table 9, the characteristics of the text are markedly different.
It is interesting to see the detrimental effect that the smart operators, which increase
bias towards semantics, have on the metre. Unlike the metrically perfect limerick in
Table 9, this text requires several edit operations (two insertions and two deletions).
(Even Belloc’s original poem contains similar rhythmic imperfections – see Figure 8).
However, the text in Table 10 conveys Starget much better than that in Table 9. The only
aspect that it fails to convey is the fact that the waist is very small. Three of the unmatched
candidate literals are actually semantically accurate duplicates of the literals mapped in the
proper matches, and not extraneous semantics. Unfortunately, our mapping algorithm
does not account for these in a satisfactory manner, just as it wrongly penalised the
solution in Table 9 for the repetition of the first line.
We believe that this resulting text is comparable to the results obtained from our chart
generation system, where we adopted an exhaustive search approach to poetry generation
(Section 2.4).

7. Summary
We have presented a model of poetry generation as stochastic search, where a goal state is
a text that satisfies the constraints of meaningfulness, grammaticality and poeticness.
 62 R. Manurung et al.

Previous attempts at poetry generation (Section 2) properly addressed only a subset of

these constraints, with the exception of our initial attempt at poetry generation using chart
generation (Manurung 1999). Unfortunately, this approach was computationally very
expensive and was not very robust, i.e. if the system was unable to satisfy any of the
constraints perfectly, it would fail to produce a single solution.
We reported on McGONAGALL, our implemented instance of this model, which
operationalises the three constraints by focusing on texts that are syntactically
well-formed, meet certain pre-specified patterns of metre, and broadly convey some
pre-specified meaning. It satisfies grammaticality through the use of lexicalised
tree-adjoining grammar (LTAG) and optimises meaningfulness and poeticness by
maximising evaluation functions that measure the degree of isomorphism between the
phenotypic features exhibited by a candidate solution and the target semantics and metre.
Experiments showed that it managed to find optimal solutions for moderately sized target
Downloaded by [McMaster University] at 11:49 28 December 2014

semantics and metre forms in isolation. However, it encountered difficulties when

simultaneously considering both evaluation functions.
There is still a lot of interesting research to be carried out in this area, particularly in
exploring the different ways in which grammaticality, meaningfulness and poeticness can
be defined. The notion of grammaticality can be augmented through the use of figurative
language rules (Quinn 1982). With regards to meaningfulness, we note that our measure of
isomorphism to a target semantics actually embodies the narrower notion of faithfulness to
a specific meaning rather than the general concept of meaningfulness proposed in
Section 3.3. A more ambitious account could consider aspects such as coherence, relevance
and interestingness. As for poeticness, while there are other phonetic features such as
rhyme and alliteration that can be explored, there are many more features to be
considered, such as figurative language and metaphor.

Acknowledgements
The work described was carried out while the authors were at the School of Informatics, University
of Edinburgh. We would also like to thank members of the University of Aberdeen NLG group and
the anonymous reviewers for their comments in preparing this article.

Notes
1. According to the Oxford English Dictionary, poetic license is the ‘deviation from recognised
form or rule, indulged in by a writer or artist for the sake of effect’.
2. A Gorn address is a list of integers that indicates the position of a node within a tree. The Gorn
address of the i-th child of node n is the Gorn address of n with i appended to the end of the list.
The Gorn address of a root node is , or sometimes 0 for convenience.

References

Aamodt, A., and Plaza, E. (1994), ‘Case-based Reasoning: Foundational Issues, Methodological
Variations, and System Approaches’, AI Communications, 7, 39–59.
Angeline, P.J. (1996), ‘Genetic Programming’s Continued Evolution’, in Advances in Genetic
Programming (Vol 2), eds. P.J. Angeline and K.E. Kinnear, Cambridge, USA: MIT Press,
pp. 89–110.
Attridge, D. (1995), Poetic Rhythm: An Introduction, Cambridge, UK: Cambridge University Press.
 Journal of Experimental & Theoretical Artificial Intelligence 63

Bäck, T., Fogel, D., and Michalewicz, Z. (Eds.) (1997), Handbook of Evolutionary Computation, New
York: Oxford University Press and Bristol: Institute of Physics Publishing.
Bailey, R.W. (1974), ‘Computer-assisted Poetry: the Writing Machine Is for Everybody’,
in Computers in the Humanities, ed. J.L. Mitchell, Edinburgh, UK: Edinburgh University
Press, pp. 283–295.
Baker, J.E. (1987), ‘Reducing Bias and Inefficiency in the Selection Algorithm’, in Proceedings of the
Second International Conference on Genetic Algorithms, ed. J.J. Grefenstette, Cambridge,
USA: Lawrence Erlbaum Associates, pp. 14–21.
Belloc, J.H.P. (1991), The Bad Child’s Book of Beasts, London, UK: Jonathan Cape.
Boden, M.A. (1990), The Creative Mind: Myths and Mechanisms, London, UK: Weidenfeld and
Nicolson.
Cheng, H. 2002, Modelling Aggregation Motivated Interactions in Descriptive Text Generation,
Ph.D. thesis, University of Edinburgh, Division of Informatics.
Davis, L., and Steenstrup, M. (1987), ‘Genetic Algorithms and Simulated Annealing: An Overview’,
Downloaded by [McMaster University] at 11:49 28 December 2014

in Genetic Algorithms and Simulated Annealing, ed. L. Davis, London, UK: Pitman, pp. 1–11.
Diaz-Agudo, B., Gervás, P., and González-Calero, P. (2002), ‘Poetry generation in COLIBRI’, in
Proceedings of the 6th European Conference on Case Based Reasoning (ECCBR 2002),
Aberdeen, UK, pp. 73–102.
Donald, B.R. (n.d.), ‘ALFRED the Mail Agent’, http://www.cs.cornell.edu/home/brd/poem.html.
Falkenhainer, B., Forbus, K.D., and Gentner, D. (1989), ‘The Structure-mapping Engine: Algorithm
and Examples’, Artificial Intelligence, 41, 1–63.
Gervás, P. (2000), ‘WaSP: Evaluation of Different Strategies for the Automatic Generation of
Spanish Verse’, in Proceedings of the AISB’00 Symposium on Creative and Cultural Aspects and
Applications of AI and Cognitive Science, Birmingham, UK: AISB, pp. 93–100.
Gervás, P. (2002), ‘Exploring Quantitative Evaluations of the Creativity of Automatic Poets’,
in Proceedings of the 2nd. Workshop on Creative Systems, Approaches to Creativity in Artificial
Intelligence and Cognitive Science, 15th European Conference on Artificial Intelligence (ECAI
2002), Lyon, France, pp. 39–46.
Goldberg, D.E. (1989), Genetic Algorithms in Search, Optimisation, and Machine Learning, Reading,
USA: Addison-Wesley.
Hartman, C.O. (1996), Virtual Muse: Experiments in Computer Poetry, Middletown, CT: Wesleyan
University Press.
Hobbs, J. (1985), ‘Ontological Promiscuity’, in Proceedings of the 23rd Annual Meeting of the
Association for Computational Linguistics, Chicago, USA: The Association for Computational
Linguistics, pp. 61–69.
Joshi, A.K. (1987), ‘The Relevance of Tree Adjoining Grammars to Generation’, in Natural
Language Generation: New Results in Artificial Intelligence, ed. G. Kempen, Dordrecht,
The Netherlands: Martinus Nijhoff Press, pp. 233–252.
Jurafsky, D.S., and Martin, J.H. (2000), Speech and Language Processing: an Introduction to
Natural Language Processing, Computational Linguistics, and Speech Recognition, Upper
Saddle River, NJ: Prentice-Hall.
Kallmeyer, L. (2002), ‘Using an Enriched TAG Derivation Structure As Basis for Semantics’, in
Proceedings of the Sixth International Workshop on Tree Adjoining Grammar and Related
Frameworks (TAGþ6), Universita di Venezia, Venezia, pp. 101–110.
Kay, M. (1996), ‘Chart Generation’, in Proceedings of the 34th Annual Meeting of the Association for
Computational Linguistics, Santa Cruz, USA: ACL, pp. 200–204.
Koller, A., and Striegnitz, K. (2002), ‘Generation As Dependency Parsing’, in Proceedings of the
40th Anniversary Meeting of the Association for Computational Linguistics, Philadelphia, USA,
pp. 17–24.
Kurzweil, R. (2001), ‘Ray Kurzweil’s Cybernetic Poet’, http://www.kurzweilcyberart.com/poetry.
Levin, S.R. (1962), ‘Linguistic Structures in Poetry’, (Number 23), in Janua Linguarum,
Mouton, Gravenhage.
 64 R. Manurung et al.

Love, B.C. (2000), ‘A Computational Level Theory of Similarity’, in Proceedings of the 22nd Annual
Meeting of the Cognitive Science Society, Philadelphia, USA, pp. 316–321.
Manurung, H.M. (1999), ‘A Chart Generator for Rhythm Patterned Text’, in Proceedings of the
First International Workshop on Literature in Cognition and Computer, Tokyo, Japan,
pp. 15–19.
Manurung, H. M. (2003), ‘An Evolutionary Algorithm Approach to Poetry Generation’, PhD thesis,
University of Edinburgh, School of Informatics.
Meteer, M. (1991), ‘Bridging the Generation Gap Between Text Planning and Linguistic
Realisation’, Computational Intelligence, 7, 296–304.
Mitchell, M. (1996), An Introduction to Genetic Algorithms, Cambridge, USA: MIT Press.
Nicolov, N., Mellish, C., and Ritchie, G. (1995), ‘Sentence Generation from Conceptual Graphs’, in
Proceedings of the International Conference on Conceptual Structures, (Number 954), in,
Springer-Verlag, Santa Cruz, USA Lecture Notes in Artificial Intelligence.
Quinn, A. (1982), Figures of Speech: 60 Ways to Turn a Phrase, Davis, CA: Hermagoras Press.
Reiter, E., and Dale, R. (2000), Building Natural Language Generation Systems (1st ed.), Cambridge,
Downloaded by [McMaster University] at 11:49 28 December 2014

UK: Cambridge University Press.

Sankoff, D. & Kruskal, J. B., (Eds.) (1983), Time warps, string edits and macromolecules: The theory
and practice of sequence comparison, Reading, USA: Addison-Wesley.
Schabes, Y. (1990), ‘Mathematical and Computational Aspects of Lexicalized Grammars’, PhD
thesis, University of Pennsylvania, Computer Science Department.
Stone, M., and Doran, C. (1997), ‘Sentence Planning as Description using Tree Adjoining
Grammar’, in Proceedings of the 35th Annual Meeting of the Association for Computational
Linguistics, Madrid, Spain: The Association for Computational Linguistics, pp. 198–205.
Vijay-Shanker, K., and Joshi, A.K. (1988), ‘Feature Structure Based Tree Adjoining Grammars’,
in Proceedings of 12th International Conference of Computational Linguistics, Budapest,
Hungary, pp. 714–720.
Weide, R.L. (1996), ‘Carnegie Mellon University pronouncing dictionary’. http://
www.speech.cs.cmu.edu/cgi-bin/cmudict.